AMD AI Workbench datasets fine-tuning

Datasets

AMD AI Workbench allows you to create and manage datasets on the platform for fine-tuning purposes. Datasets are collections of data points, where each data point represents a conversation, with or without the model’s expected response, depending on the use case.

Uploading a dataset

Navigate to the Datasets page to upload a dataset. You can provide a name and description for the dataset, and upload a file in JSONL format.

The type of dataset should be selected when uploading the file, and the format of each row should correspond to the selected dataset type.

Data formats

Depending on the selected dataset type, each row in the JSONL file should be formatted as shown below.

Note

JSONL formatting In a real JSONL file, each object must be on a single line. Here, objects are formatted for readability.

Fine-tuning

{
  "messages": [
    {
      "role": "system",
      "content": "You are a helpful Geography bot. Only answer questions about Geography."
    },
    { "role": "user", "content": "What is the capital of France?" },
    { "role": "assistant", "content": "Paris" }
  ]
}

Direct preference

{
  "prompt_messages": [
    {
      "role": "system",
      "content": "You are a helpful Geography bot. Only answer questions about Geography."
    },
    { "role": "user", "content": "What is the capital of France?" }
  ],
  "chosen_messages": [{ "role": "assistant", "content": "Paris" }],
  "rejected_messages": [{ "role": "assistant", "content": "I do not know" }]
}

Unarchiving datasets

If you have identified a dataset that is no longer relevant to your use cases, you can choose to delete. This happens by selecting the dataset(s) and clicking the “Delete” button in the actions menu.